The AccelMR 2020 Challenge invites researchers from the medical imaging community to submit methods that define the non-linear mapping between pairs of magnetic resonance images (MRI) acquired at multiple resolutions. The scope of the challenge is to identify the most accurate method that can reconstruct a high-resolution MR image that is of diagnostic quality from image data acquired at low resolution, thus faster. Faster MRI systems could potentially push aside X-ray and CT for some applications and avoid exposure to ionizing radiation. Furthermore, this could also potentially help in eliminating sedation or general anesthesia needed for MRI scanning of young children to make them stay still during the acquisition process.
Although acceleration of MR acquisition through post-reconstruction deep-learning techniques has seen a lot of interest recently in the medical imaging community (Hyun et al. 2018; Sandino and Cheng 2017), recent notable attempts in this area has been made through a joint effort between NYU’s Center for Advanced Imaging Innovation and Research (CAI2R), and the Facebook Artificial Intelligence Research group (https://www.aiin.healthcare/sponsored/9667/topics/artificial-intelligence/nyus-daniel-sodickson-ai-facebook-and-why-faster-mr?fbclid=IwAR23x92NHauyLiJj5PpH_dcNKFgYlRy1JPTXD-GXmSZXNh72hzNuniDY5ZY). The mainstay of these approaches is to train artificial neural networks that can recognize the underlying structure of the images to fill in views omitted from the accelerated scan. However, currently no publicly available datasets exist of higher-/low- resolution image pairs that can be used for training purposes in these approaches. Therefore, existing methods rely mainly on the paired-data simulated through either adding noise or downsampling. The organizers of AccelMR try to overcome this shortcoming in the medical imaging community. After the approval of the Institutional Review Board (IRB; Pro00010785), multi-sequence MRI data was collected at multiple resolutions from adult subjects, imaging the same field-of-view (FOV). The challenge organizers will provide 45 3D T1- and T2-weighted sequences of brain MRI acquired at three different resolutions (high/diagnostic quality, half, and one-quarter) using the 3T Discovery MR750 scanner by General Electric (Waukesha, WI) for training. Contestants will be ranked based on method performance on an independent test dataset consisting of 5 subjects. On each data set, we will compute the following evaluation metrics: Peak Signal-to-Noise Ratio, Structural Similarity Index between the predicted/reconstructed image (separately from half- and quarter-resolution), and the original high-resolution image.